FROM GOOGLE'S DISCOVERY TO A DIFFERENT AGENTIC ARCHITECTURE : The exploit_agent Case, the Risks of Unrestricted Autonomous Cyber Operations, and the ST-T1024 A1–A16 Framework
On November 16, 2026, Google's Threat Intelligence Group (GTIG), working with Mandiant, published findings that mark an important development in the operational use of artificial intelligence in cyber operations. https://x.com/i/status/2123457542345179163 GTIG identified an exposed staging server hosting a multi-agent framework calledexploit_agent. The framework integrated multiple frontier and locally hosted largelanguage models and was configured to continuously analyse software patches, identify potential vulnerabilities, draft exploits and execute them automatically, without human intervention at those stages. This finding moves the discussion of agentic artificial intelligence beyond theoreticalconcerns. The central issue is no longer simply whether AI can generate sophisticated code or assist ahuman attacker. It is whether an AI agent can autonomously transform an objective into asequence of consequential operations, interact with external systems, evaluate executionresults and continue acting at machine speed. This paper examines that development and proposes a different architectural approach:ST-T1024 Agentic Architecture, documented through A1–A16. Its fundamental principle is that an agent may exercise autonomy within a human-definedoperational perimeter, without acquiring authority to redefine that perimeter. The human defines the knowledge library, permissible goals, actions, permissions, execution constraints and verification requirements. The agent operates within those specifications. The proposition is not that autonomous execution must be eliminated. It is that autonomous execution must be separated from autonomous authority. THE AGENT MAY EXECUTE. IT MUST NOT BECOME THE AUTHOR OF ITS OWNAUTHORITY.
Authors
- Team LLM-DIPLOMAT
- M. Taalabi
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-20
- DOI
- https://doi.org/10.5281/zenodo.24386362
- Primary Topic
- Information and Cyber Security
- Type
- article
- Field-Weighted Citation Impact
- 0.00